I open-sourced GOD HAND: a local-first tool for exploring tabular data

SATNAM — Wed, 11 Nov 2026 23:30:21 +0000

Today I'm releasing GOD HAND OF STATISTICS as open source under the Apache-2.0 licence.

👉 github.com/ProbotisOP/GodHandOFStatistics

What it is

A local web app for exploring tabular data. You drop in files, and it profiles, cleans and connects them, searches for patterns, and shows only the patterns that survive statistical checks.

It runs entirely on your machine (server bound to 127.0.0.1). No account, no API key, no cloud.

Why

Automated exploration tests hundreds of hypotheses, and some will "pass" by chance. On a file of pure random numbers:

$ uv run python examples/noise_challenge.py
  naive testing, p < 0.05 one at a time : 10 'significant' of 230 tests  <- all false alarms
  GOD HAND, 231 hypotheses on the ledger : 0 patterns reported

GOD HAND is built so that the second line is the normal outcome.

Highlights

Statistics you can trust

Every hypothesis is registered on a ledger before it is tested; the ledger is the only source of p-values
Benjamini–Hochberg false-discovery control across the whole search
Sealed 30% holdout (≥ 2,000 rows) or cross-fitting (300–2,000 rows), with the candidate list frozen first
Permutation tests (Besag–Clifford), effect sizes with intervals (Spearman ρ, Cliff's δ), design-effect correction for repeated rows
Evidence grades A–E, explained in plain language without causal claims

Finds what matters

Relationships, group differences, segments, interactions, Simpson's paradox, change points, differences between files
Exact formula detection, e.g. Revenue = Avg_Spend × Daily_Customers × 0.03
Near-copy columns merged, so findings aren't duplicated

Works with your files

CSV, TSV, Excel, Parquet, JSON
Microsoft Access (.mdb, .accdb), SQLite, DuckDB, dBase, all read-only and detected by their file signature
Multi-file joins and appends, with warnings for row inflation and unit mismatches

Clean and export

Green / yellow / red health light per table
One-click fixes, exported as cleaned CSV/Parquet plus a Python (polars) or SQL recipe

Explore

Interactive Vega-Lite charts: zoom, pan, hover, isolate a series, export PNG/SVG/CSV
A three.js Data Universe: columns are stars, findings are beams coloured by direction, and clicking shows the evidence

Stack

Python 3.12–3.13 (polars, numpy, scipy, scikit-learn, FastAPI) · React 19 + TypeScript + Vite · three.js · Vega-Lite · tested with pytest, Hypothesis and Playwright.

Get started

git clone https://github.com/ProbotisOP/GodHandOFStatistics && cd GodHandOFStatistics
uv sync
(cd apps/web && npm install && npm run build)
uv run python -m godhand.server        # http://127.0.0.1:8000

Requires uv and Node.js.

Status

This is release 0.1. Known limits: runs are kept in memory, very large scatter plots are slow, and password-protected Access files aren't supported.

Coming next: drift watch (re-test findings on new data), a local AI copilot answering only from the ledger, and persistent runs.

Get involved

Try it on a real dataset and tell me what it found, what it missed, and where it got the statistics wrong. Issues, ideas and pull requests are welcome. See CONTRIBUTING.md.

⭐ If it's useful, a star helps others find it: github.com/ProbotisOP/GodHandOFStatistics

DEV Community: SATNAM